Variance Reduction for Policy Gradient...

Post on 11-Jul-2020

0 views 0 download

Transcript of Variance Reduction for Policy Gradient...

Variance Reduction for Policy Gradient Methods

March 13, 2017

Reward Shaping

Reward Shaping

Reward Shaping

I Reward shaping: r(s, a, s ′) = r(s, a, s ′) + γΦ(s ′)− Φ(s) for arbitrary“potential” Φ

I Theorem: r admits the same optimal policies as r .1

I Proof sketch: suppose Q∗ satisfies Bellman equation (T Q = Q). If wetransform r → r , policy’s value function satisfies Q(s, a) = Q∗(s, a)− Φ(s)

I Q∗ satisfies Bellman equation ⇒ Q also satisfies Bellman equation

1A. Y. Ng, D. Harada, and S. Russell. “Policy invariance under reward transformations: Theory and application to reward shaping”. ICML. 1999.

Reward ShapingI Theorem: R admits the same optimal policies as R. A. Y. Ng, D. Harada, and S. Russell.

“Policy invariance under reward transformations: Theory and application to rewardshaping”. ICML. 1999

I Alternative proof: advantage function is invariant. Let’s look at effect on V π and Qπ:

E[r0 + γr1 + γ2r2 + . . .

]condition on either s0 or (s0, a0)

= E[r0 + γ r1 + γ2r2 + . . .

]= E

[(r0 + γΦ(s1)− Φ(s0)) + γ(r1 + γΦ(s2)− Φ(s1)) + γ2(r2 + γΦ(s3)− Φ(s2)) + . . .

]= E

[r0 + γr1 + γ2r2 + · · · − Φ(s0)

]Thus,

V π(s) = V π(s)− Φ(s)

Qπ(s) = Qπ(s, a)− Φ(s)

Aπ(s) = Aπ(s, a)

Aπ(s, π(s)) = 0 at all states ⇒ π is optimal

Reward Shaping and Problem Difficulty

I Shape with Φ = V ∗ ⇒ problem is solved in one step of value iteration

I Shaping leaves policy gradient invariant (and just adds baseline to estimator)

E[∇θ log πθ(a0 | s0)(r0 + γΦ(s1)− Φ(s0)) + γ(r1 + γΦ(s2)− Φ(s1))

+ γ2(r2 + γΦ(s3)− Φ(s2)) + . . .]

= E[∇θ log πθ(a0 | s0)(r0 + γr1 + γ2r2 + · · · − Φ(s0))

]= E

[∇θ log πθ(a0 | s0)(r0 + γr1 + γ2r2 + . . . )

]

Reward Shaping and Policy GradientsI First note connection between shaped reward and advantage function:

Es1 [r0 + γV π(s1)− V π(s0) | s0 = s, a0 = a] = Aπ(s, a)

Now considering the policy gradient and ignoring all but first shaped reward (i.e., pretendγ = 0 after shaping) we get

E

[∑t

∇θ log πθ(at | st)rt

]= E

[∑t

∇θ log πθ(at | st)(rt + γV π(st+1)− V π(st))

]

= E

[∑t

∇ log πθ(at | st)Aπ(st , at)

]I Compromise: use more aggressive discount γλ, with λ ∈ (0, 1): called generalized

advantage estimation

E

[∑t

∇θ log πθ(at | st)∞∑k=0

(γλ)k rt+k

]

Reward Shaping – Summary

I Reward shaping transformation leaves policy gradient and optimal policyinvariant

I Shaping with Φ ≈ V π makes consequences of actions more immediate

I Shaping, and then ignoring all but first term, gives policy gradient

[Aside] Reward Shaping: Very Important in PracticeI I. Mordatch, E. Todorov, and Z. Popovic. “Discovery of complex behaviors through

contact-invariant optimization”. ACM Transactions on Graphics (TOG) 31.4 (2012),p. 43

I Y. Tassa, T. Erez, and E. Todorov. “Synthesis and stabilization of complex behaviorsthrough online trajectory optimization”. Intelligent Robots and Systems (IROS), 2012IEEE/RSJ International Conference on. IEEE. 2012, pp. 4906–4913

Variance Reduction for Policy Gradients

Variance Reduction

I We have the following policy gradient formula:

∇θEτ [R] = Eτ

[T−1∑t=0

∇θ log π(at | st , θ)Aπ(st , at)

]

I Aπ is not known, but we can plug in At , an advantage estimator

I Previously, we showed that taking

At = rt + rt+1 + rt+2 + · · · − b(st)

for any function b(st), gives an unbiased policy gradient estimator.b(st) ≈ V π(st) gives variance reduction.

The Delayed Reward Problem

I With policy gradient methods, we are confounding the effect of multipleactions:

At = rt + rt+1 + rt+2 + · · · − b(st)

mixes effect of at , at+1, at+2, . . .

I SNR of At scales roughly as 1/TI Only at contributes to signal Aπ(st , at), but at+1, at+2, . . . contribute to

noise.

Variance Reduction with DiscountsI Discount factor γ, 0 < γ < 1, downweights the effect of rewards that are far

in the future—ignore long term dependencies

I We can form an advantage estimator using the discounted return:

Aγt = rt + γrt+1 + γ2rt+2 + . . .︸ ︷︷ ︸discounted return

−b(st)

reduces to our previous estimator when γ = 1.

I So advantage has expectation zero, we should fit baseline to be discountedvalue function

V π,γ(s) = Eτ[r0 + γr1 + γ2r2 + . . . | s0 = s

]I Discount γ is similar to using a horizon of 1/(1− γ) timesteps

I Aγt is a biased estimator of the advantage function

Value Functions in the Future

I Baseline accounts for and removes the effect of past actions

I Can also use the value function to estimate future rewards

rt + γV (st+1) cut off at one timestep

rt + γrt+1 + γ2V (st+2) cut off at two timesteps

. . .

rt + γrt+1 + γ2rt+2 + . . . ∞ timesteps (no V )

Value Functions in the Future

I Subtracting out baselines, we get advantage estimators

A(1)t = rt + γV (st+1)−V (st)

A(2)t = rt + rt+1 + γ2V (st+2)−V (st)

. . .

A(∞)t = rt + γrt+1 + γ2rt+2 + . . .−V (st)

I A(1)t has low variance but high bias, A

(∞)t has high variance but low bias.

I Using intermediate k (say, 20) gives an intermediate amount of bias and variance

Finite-Horizon Methods: Advantage Actor-Critic

I A2C / A3C uses this fixed-horizon advantage estimator

I Pseudocode

for iteration=1, 2, . . . doAgent acts for T timesteps (e.g., T = 20),For each timestep t, compute

Rt = rt + γrt+1 + · · ·+ γT−t+1rT−1 + γT−tV (st)

At = Rt − V (st)

Rt is target value function, in regression problemAt is estimated advantage function

Compute loss gradient g = ∇θ

∑Tt=1

[− log πθ(at | st)At + c(V (s)− Rt)

2]

g is plugged into a stochastic gradient descent variant, e.g., Adam.end for

V. Mnih, A. P. Badia, M. Mirza, A. Graves, T. P. Lillicrap, et al. “Asynchronous methods for deep reinforcement learning”. (2016)

A3C Video

A3C Results

0 2 4 6 8 10 12 14Training time (hours)

0

2000

4000

6000

8000

10000

12000

14000

16000

Sco

re

Beamrider

DQN1-step Q1-step SARSAn-step QA3C

0 2 4 6 8 10 12 14Training time (hours)

0

100

200

300

400

500

600

Sco

re

Breakout

DQN1-step Q1-step SARSAn-step QA3C

0 2 4 6 8 10 12 14Training time (hours)

30

20

10

0

10

20

30

Sco

re

Pong

DQN1-step Q1-step SARSAn-step QA3C

0 2 4 6 8 10 12 14Training time (hours)

0

2000

4000

6000

8000

10000

12000

Sco

re

Q*bert

DQN1-step Q1-step SARSAn-step QA3C

0 2 4 6 8 10 12 14Training time (hours)

0

200

400

600

800

1000

1200

1400

1600

Sco

re

Space Invaders

DQN1-step Q1-step SARSAn-step QA3C

Reward Shaping

TD(λ) Methods: Generalized Advantage EstimationI Recall, finite-horizon advantage estimators

A(k)t = rt + γrt+1 + · · ·+ γk−1rt+k−1 + γkV (st+k)− V (st)

I Define the TD error δt = rt + γV (st+1)− V (st)

I By a telescoping sum,

A(k)t = δt + γδt+1 + · · ·+ γk−1δt+k−1

I Take exponentially weighted average of finite-horizon estimators:

Aλ = A(1)t + λA

(2)t + λ2A

(3)t + . . .

I We obtain

Aλt = δt + (γλ)δt+1 + (γλ)2δt+2 + . . .

I This scheme named generalized advantage estimation (GAE) in [1], though versions haveappeared earlier, e.g., [2]. Related to TD(λ)

J. Schulman, P. Moritz, S. Levine, M. Jordan, and P. Abbeel. “High-dimensional continuous control using generalized advantage estimation”.(2015)

H. Kimura and S. Kobayashi. “An Analysis of Actor/Critic Algorithms Using Eligibility Traces: Reinforcement Learning with Imperfect ValueFunction.” ICML. 1998

Choosing parameters γ, λ

Performance as γ, λ are varied

0 100 200 300 400 500

number of policy iterations

2.5

2.0

1.5

1.0

0.5

0.0

cost

3D Biped

γ=0.96,λ=0.96

γ=0.98,λ=0.96

γ=0.99,λ=0.96

γ=0.995,λ=0.92

γ=0.995,λ=0.96

γ=0.995,λ=0.98

γ=0.995,λ=0.99

γ=0.995,λ=1.0

γ=1,λ=0.96

γ=1, No value fn

TRPO+GAE Video